RUCF for your AI assistant

1. What it does

The skill is a folder of instructions, reference files and one small Python script. Your assistant reads it when you ask for a UX audit and follows the framework step by step, the same way a trained practitioner would.

It collects behaviour (what users do) and perception (what users say) in separate passes and maps both onto the twelve signals of the scoring model. It tags each value with its evidence class, runs the arithmetic in code and places each of the four surfaces on the say-do matrix. Then it writes a report in a fixed twelve-section structure, so two runs on the same product can be compared line by line.

The script is the same engine as the scorecard on this site, ported to Python. Before each release it is checked against the website engine on 236 test cases, and the version you download matched all of them. Wherever your assistant can run Python, the numbers come out the same in Claude, ChatGPT, Gemini and Grok.

2. What it cannot do

An AI assistant cannot watch your users. With only a URL it can walk your public pages, run an accessibility scan and read public reviews, but its view of how hard the product is to use is still its own judgment. RUCF weights expert judgment at 0.30, and that applies to an AI the same as to a person (why the weight is that low).

So a URL-only run comes back with a low confidence figure, typically between 0.30 and 0.45, and a list of things to measure. It names candidate silent friction, and then it tells you which event to log or which question to ask to confirm or rule each one out. That list is useful, but it is a plan, and the skill labels it as one.

Two more limits. Below a few hundred active users the framework should not be applied, and the skill will say so. And no run can give you a real tenth percentile unless you supply behaviour data broken down by user or segment.

3. Three modes

The assistant picks the mode from what you give it, and names it at the top of the report.

In a data audit, the assistant's own reading of reviews or tickets enters at 0.50. If you check a sample of 20 or more of its codes and agree with at least 80%, that evidence moves up to 0.70. The skill offers you the sample.

4. Install

Anthropic published the skill format, a folder with a SKILL.md file, as an open standard in December 2025, and the other major assistants adopted it during 2026. One zip works in all of them. Menus change often. If a step below does not match what you see, the rule is: the zip goes wherever your assistant accepts skills. If your assistant has no skills on your plan, paste SKILL.md into its custom instructions and attach the other files.

Claude

1. Turn on code execution in Settings, under Capabilities.
2. In the same place, find Skills and upload rucf-audit.zip.
3. In Claude Code, unzip it into ~/.claude/skills/ instead.

ChatGPT

1. If your plan shows Skills, upload rucf-audit.zip there.
2. If not, create a custom GPT. Paste the contents of SKILL.md into its instructions, upload the files from references/, scripts/ and examples/ as knowledge, and turn on code interpreter so the script can run.

Gemini

1. Skills began rolling out in the Gemini app in October 2026. If you have them, add rucf-audit.zip.
2. If not, create a Gem with SKILL.md as its instructions and the reference files attached. A Gem may not be able to run the script. In that case the skill tells it to do the arithmetic by hand and show every step, so you can check it.

Grok

1. Open Skills and import rucf-audit.zip. Grok accepts .zip files directly.

Coding agents

Codex, Cursor, GitHub Copilot, Gemini CLI and other agents that read SKILL.md folders take the unzipped rucf-audit folder in their skills directory. Check your agent's documentation for the path.

5. How to ask

Ask for an RUCF audit by name. Asked only for a "UX audit", most assistants will answer from their own habits without opening the skill. Then say what the product is for and give it whatever evidence you have. Replace the parts in brackets.

Desk review

Use the rucf-audit skill. Run a desk review of [https://your-site.com].
The main thing people come to do is [book a table].
Product type: [ecommerce].

Guided scorecard

Use the rucf-audit skill. Take me through the scorecard
questions for [product], behaviour first, then perception.

Data audit

Use the rucf-audit skill. Run a data audit of [product].
Attached: [our Q3 funnel export, the last 100 support tickets,
our app store reviews]. The main task is [sending an invoice].
About [4,000] monthly active users. [2] releases since this data.

The assistant asks for anything it cannot find itself: product type, the main task and what counts as success, active users, and how many releases have shipped since the evidence was gathered. Releases matter because RUCF lowers confidence in evidence each time the measured surface ships.

6. The report

Every report has the same twelve sections, in the same order:

1. Run header, with the composite written as score and confidence, for example 2.8 @ 0.75
2. Intent: the job users come to do, and what counts as success
3. Evidence inventory: every source, its class, weight and age
4. Surface table: behaviour, perception, the gap and the quadrant for each surface
5. Diagnosis by surface, with the mechanism and user quotes
6. Silent friction candidates, each with the measurement that would confirm it
7. Top three treatments, each with a success criterion written in advance
8. One named strength to protect
9. Accessibility gate against WCAG 2.2 AA, kept out of the composite
10. Tail: real percentiles from segmented data, or a risk flag
11. Instrument next: the events and questions that would raise confidence most
12. What the report cannot tell you

The download includes a complete worked example for a fictional invoicing app. It found silent friction on the Task surface (behaviour 33, perception 83) and phantom friction on Belief (behaviour 83, perception 25). Those two surfaces need opposite treatments, and an averaged score would have reported both as 58.

7. Check the download

Public skill directories have carried malicious skills, so check before you install anything, this file included. The script makes no network calls and imports only the Python standard library. Read it first if you like: it is one file, scripts/rucf_score.py.

SHA-256 of rucf-audit.zip v1.0.0:

d10c854f2285b48ac451255f20faa308eb06b0d36978b59cbca07375169316f1

On a Mac or Linux, shasum -a 256 rucf-audit.zip should print the same string. After unzipping, this command runs the 236 test cases against your copy:

python3 rucf-audit/scripts/rucf_score.py selftest

8. Your data

The skill runs inside your own assistant. Nothing comes back to this site, and nothing is stored by RUCF. Your assistant provider's terms still apply to whatever you upload, so remove names, email addresses and account IDs from exports and tickets first. The audit does not need them.

In a desk review the skill tells the assistant to read public pages only. It does not sign up, log in, submit forms, start a payment or contact anyone.

Download rucf-audit.zip

---

Related: The scorecard · The scoring model · Silent friction · Contact